From 074b985f3855193bb47fb4055abb6b12f09f48d7 Mon Sep 17 00:00:00 2001 From: Pablo Marquez Tello Date: Fri, 13 Oct 2023 10:03:58 +0100 Subject: FuseBatchNorm changes to enable fp16 in armv8a multi_isa builds * FP16 kernels must be instantiated in fp16.cpp. * Partially resolves MLCE-1102 Change-Id: Ie652203876a0ac12b025e96d20990b6efb21e772 Signed-off-by: Pablo Marquez Tello Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10477 Tested-by: Arm Jenkins Benchmark: Arm Jenkins Reviewed-by: Jakub Sujak Comments-Addressed: Arm Jenkins --- .../fuse_batch_normalization/generic/fp16.cpp | 17 ++- .../fuse_batch_normalization/generic/impl.h | 120 ++++++++++++++++++- .../kernels/fuse_batch_normalization/nchw/all.cpp | 133 +-------------------- 3 files changed, 134 insertions(+), 136 deletions(-) diff --git a/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp b/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp index 2821af32ce..8f47ecba8f 100644 --- a/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/generic/fp16.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,6 +43,21 @@ void fused_batch_normalization_conv_f16(const ITensor *conv_weights, return fused_batch_normalization_conv(conv_weights, conv_bias, fused_weights, fused_bias, bn_mean, bn_var, bn_beta, bn_gamma, epsilon, window); } + +void fused_batch_normalization_dwc_nchw_f16(const ITensor *dwc_weights, + const ITensor *dwc_bias, + ITensor *fused_weights, + ITensor *fused_bias, + const ITensor *bn_mean, + const ITensor *bn_var, + const ITensor *bn_beta, + const ITensor *bn_gamma, + float epsilon, + const Window &window) +{ + return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, + bn_var, bn_beta, bn_gamma, epsilon, window); +} } // namespace cpu } // namespace arm_compute #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ diff --git a/src/cpu/kernels/fuse_batch_normalization/generic/impl.h b/src/cpu/kernels/fuse_batch_normalization/generic/impl.h index 6fa843263a..d807148e37 100644 --- a/src/cpu/kernels/fuse_batch_normalization/generic/impl.h +++ b/src/cpu/kernels/fuse_batch_normalization/generic/impl.h @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H -#define SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H +#ifndef ACL_SRC_CPU_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H +#define ACL_SRC_CPU_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H #include "arm_compute/core/Helpers.h" @@ -144,6 +144,120 @@ void fused_batch_normalization_conv(const ITensor *conv_weights, }, conv_w_in, conv_w_out); } +template +void fused_batch_normalization_dwc_nchw(const ITensor *dwc_weights, + const ITensor *dwc_bias, + ITensor *fused_weights, + ITensor *fused_bias, + const ITensor *bn_mean, + const ITensor *bn_var, + const ITensor *bn_beta, + const ITensor *bn_gamma, + float epsilon, + const Window &window) +{ + using ScalarType = T; + const int size = 16 / dwc_weights->info()->element_size(); + using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; + + const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights); + const bool run_in_place_bias = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias); + + // Set build options + Window win = window; + win.set(Window::DimX, Window::Dimension(0, 1, 1)); + + const int window_step_x = size; + const auto window_start_x = static_cast(window.x().start()); + const auto window_end_x = static_cast(window.x().end()); + + Iterator dwc_w_in(dwc_weights, win); + Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win); + + const auto dwc_bias_in = + (dwc_bias != nullptr ? reinterpret_cast(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); + auto dwc_bias_out = + (run_in_place_bias ? dwc_bias_in + : reinterpret_cast(fused_bias->ptr_to_element(Coordinates(0, 0)))); + + const auto input_mean = reinterpret_cast(bn_mean->ptr_to_element(Coordinates(0, 0))); + const auto input_var = reinterpret_cast(bn_var->ptr_to_element(Coordinates(0, 0))); + const auto input_gamma = (bn_gamma != nullptr) + ? reinterpret_cast(bn_gamma->ptr_to_element(Coordinates(0, 0))) + : nullptr; + const auto input_beta = (bn_beta != nullptr) + ? reinterpret_cast(bn_beta->ptr_to_element(Coordinates(0, 0))) + : nullptr; + + auto mean_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto var_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto gamma_vec = wrapper::vdup_n(ScalarType(1), ExactTagType{}); + auto beta_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + auto rvar_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); + const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{}); + + auto mean = ScalarType(0.0); + auto var = ScalarType(0.0); + auto gamma = ScalarType(1.0); + auto beta = ScalarType(0.0); + auto dwc_bias_in_scalar = ScalarType(0.0); + execute_window_loop( + win, + [&](const Coordinates &id) + { + var = input_var[id[2]]; + if (input_gamma != nullptr) + { + gamma = input_gamma[id[2]]; + } + + if (id[1] == 0) + { + mean = input_mean[id[2]]; + + // Construct vectors + mean_vec = wrapper::vdup_n(mean, ExactTagType{}); + if (input_beta != nullptr) + { + beta = input_beta[id[2]]; + beta_vec = wrapper::vdup_n(beta, ExactTagType{}); + } + + if (dwc_bias_in != nullptr) + { + dwc_bias_in_scalar = dwc_bias_in[id[2]]; + } + + auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); + dwc_bias_out[id[2]] = (dwc_bias_tmp_scalar * gamma) + beta; + } + + int x = window_start_x; + auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); + auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); + var_vec = wrapper::vdup_n(var, ExactTagType{}); + gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); + rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); + + for (; x <= (window_end_x - window_step_x); x += window_step_x) + { + auto wn = wrapper::vloadq(dwc_w_in_ptr + x); + wn = wrapper::vmul(wn, rvar_vec); + wn = wrapper::vmul(wn, gamma_vec); + + // Store results + wrapper::vstore(dwc_w_out_ptr + x, wn); + } + + // Compute left-over elements + for (; x < window_end_x; ++x) + { + *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; + } + }, + dwc_w_in, dwc_w_out); +} + } // namespace cpu } // namespace arm_compute -#endif //SRC_CORE_NEON_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H +#endif // ACL_SRC_CPU_KERNELS_FUSE_BATCH_NORMALIZATION_GENERIC_IMPL_H diff --git a/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp b/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp index c0b0dfd4dc..25580e1bec 100644 --- a/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp +++ b/src/cpu/kernels/fuse_batch_normalization/nchw/all.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2022 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,120 +28,6 @@ namespace arm_compute { namespace cpu { -template -void fused_batch_normalization_dwc_nchw(const ITensor *dwc_weights, - const ITensor *dwc_bias, - ITensor *fused_weights, - ITensor *fused_bias, - const ITensor *bn_mean, - const ITensor *bn_var, - const ITensor *bn_beta, - const ITensor *bn_gamma, - float epsilon, - const Window &window) -{ - using ScalarType = T; - const int size = 16 / dwc_weights->info()->element_size(); - using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; - - const bool run_in_place_weights = (fused_weights == nullptr) || (fused_weights == dwc_weights); - const bool run_in_place_bias = (fused_bias == nullptr) || (dwc_bias != nullptr && fused_bias == dwc_bias); - - // Set build options - Window win = window; - win.set(Window::DimX, Window::Dimension(0, 1, 1)); - - const int window_step_x = size; - const auto window_start_x = static_cast(window.x().start()); - const auto window_end_x = static_cast(window.x().end()); - - Iterator dwc_w_in(dwc_weights, win); - Iterator dwc_w_out(run_in_place_weights ? dwc_weights : fused_weights, win); - - const auto dwc_bias_in = - (dwc_bias != nullptr ? reinterpret_cast(dwc_bias->ptr_to_element(Coordinates(0, 0))) : nullptr); - auto dwc_bias_out = - (run_in_place_bias ? dwc_bias_in - : reinterpret_cast(fused_bias->ptr_to_element(Coordinates(0, 0)))); - - const auto input_mean = reinterpret_cast(bn_mean->ptr_to_element(Coordinates(0, 0))); - const auto input_var = reinterpret_cast(bn_var->ptr_to_element(Coordinates(0, 0))); - const auto input_gamma = (bn_gamma != nullptr) - ? reinterpret_cast(bn_gamma->ptr_to_element(Coordinates(0, 0))) - : nullptr; - const auto input_beta = (bn_beta != nullptr) - ? reinterpret_cast(bn_beta->ptr_to_element(Coordinates(0, 0))) - : nullptr; - - auto mean_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto var_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto gamma_vec = wrapper::vdup_n(ScalarType(1), ExactTagType{}); - auto beta_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - auto rvar_vec = wrapper::vdup_n(ScalarType(0), ExactTagType{}); - const auto epsilon_vec = wrapper::vdup_n(ScalarType(epsilon), ExactTagType{}); - - auto mean = ScalarType(0.0); - auto var = ScalarType(0.0); - auto gamma = ScalarType(1.0); - auto beta = ScalarType(0.0); - auto dwc_bias_in_scalar = ScalarType(0.0); - execute_window_loop( - win, - [&](const Coordinates &id) - { - var = input_var[id[2]]; - if (input_gamma != nullptr) - { - gamma = input_gamma[id[2]]; - } - - if (id[1] == 0) - { - mean = input_mean[id[2]]; - - // Construct vectors - mean_vec = wrapper::vdup_n(mean, ExactTagType{}); - if (input_beta != nullptr) - { - beta = input_beta[id[2]]; - beta_vec = wrapper::vdup_n(beta, ExactTagType{}); - } - - if (dwc_bias_in != nullptr) - { - dwc_bias_in_scalar = dwc_bias_in[id[2]]; - } - - auto dwc_bias_tmp_scalar = (dwc_bias_in_scalar - mean) / std::sqrt(var + ScalarType(epsilon)); - dwc_bias_out[id[2]] = (dwc_bias_tmp_scalar * gamma) + beta; - } - - int x = window_start_x; - auto dwc_w_in_ptr = reinterpret_cast(dwc_w_in.ptr()); - auto dwc_w_out_ptr = reinterpret_cast(dwc_w_out.ptr()); - var_vec = wrapper::vdup_n(var, ExactTagType{}); - gamma_vec = wrapper::vdup_n(gamma, ExactTagType{}); - rvar_vec = wrapper::vinvsqrt(wrapper::vadd(var_vec, epsilon_vec)); - - for (; x <= (window_end_x - window_step_x); x += window_step_x) - { - auto wn = wrapper::vloadq(dwc_w_in_ptr + x); - wn = wrapper::vmul(wn, rvar_vec); - wn = wrapper::vmul(wn, gamma_vec); - - // Store results - wrapper::vstore(dwc_w_out_ptr + x, wn); - } - - // Compute left-over elements - for (; x < window_end_x; ++x) - { - *(dwc_w_out_ptr + x) = *(dwc_w_in_ptr + x) / std::sqrt(var + ScalarType(epsilon)) * gamma; - } - }, - dwc_w_in, dwc_w_out); -} - void fused_batch_normalization_dwc_nchw_f32(const ITensor *dwc_weights, const ITensor *dwc_bias, ITensor *fused_weights, @@ -157,22 +43,5 @@ void fused_batch_normalization_dwc_nchw_f32(const ITensor *dwc_weights, bn_var, bn_beta, bn_gamma, epsilon, window); } -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) -void fused_batch_normalization_dwc_nchw_f16(const ITensor *dwc_weights, - const ITensor *dwc_bias, - ITensor *fused_weights, - ITensor *fused_bias, - const ITensor *bn_mean, - const ITensor *bn_var, - const ITensor *bn_beta, - const ITensor *bn_gamma, - float epsilon, - const Window &window) -{ - return fused_batch_normalization_dwc_nchw(dwc_weights, dwc_bias, fused_weights, fused_bias, bn_mean, - bn_var, bn_beta, bn_gamma, epsilon, window); -} -#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ - } // namespace cpu } // namespace arm_compute -- cgit v1.2.1